Encuesta aplicada al personal de MEDICVALLE

The pair wise comparison of participation effects between matched and unmatched groups in the previous section suggests that matching reduces a significant part of bias due to differences prior to entry. However, it does not account for how participating households would have fared in subsequent years if they had not participated. Such effect is captured by the composite counterfactual estimated according to methods provided in section 6.3.1 and 6.3.2 (detail calculations are given in appendix 6A). Main results are given in table 6.3. For comparison purposes, matched pair wise effects are also provided for each outcome measurement period along with the composite effects. The columns in table 6.3 provide these

comparative results for each group of participant (t=1,…, 4) and in each outcome

measurement period (v=1,..,4). Note that standard errors are not given for composite results because they are calculated from several matching results according to methods in (2) and (4) (see appendix 6A).

There are two important findings in this exercise. First, regarding the main causal question of comparing the effect of early versus late participation, the composite counterfactual results suggest that early participants have consistently fared better than late participants. Specifically, after accounting for both initial differences and potential future changes in the composition of participants and their controls, long-term participants have enjoyed relatively higher average annual consumption than short-term participants. In table 6.3, this can be seen by comparing the composite effects for each new participant group (i.e., at each t) against its preceding participant column wise. Note that the composite effect, for the most part, declines going from top (early participants) to bottom (late participants) in each column. One reason is that the effect of borrowing lasts longer than the specific period it refers to and that long- rather than short-standing participants are more likely to enjoy higher effects in terms of capacity to smooth consumption over time. Another is since participation is state dependent, at least, in this data set (see results in chapter three); the chances of repeat participation and hence further increases in consumption are higher for early than late participants.

Second, in contrast to simple pair wise effects, the composite effects provide conservative results in all comparisons except for the initial year1. This is because the composite effects take future potential counterfactuals into account whereas the pair wise estimates do not. In other words, not accounting for future potential counterfactuals overestimates impact. This is so because not participating in any earlier year does not preclude the possibility of participating in any later year and given positive effects of participation, not accounting for these chances of later participation overestimates impact of

1 _{Note that in each initial year, the pair wise effect is the same as the composite effect because t=v and} there is no need to account for future potential counterfactual.

early participation. This can be elaborated using the most early participants (i.e., t=1) whose outcome is measured in four period. The composite effect in the last outcome measurement period (v=4) takes into account the fact that some of their matched controls (i.e., non- participants at t=1) have been able to participate at t=2, t=3 or at t=4. This reduces the average effect from ETB 1488 to ETB 1238. Clearly, the difference is the counterfactual for early participants had they not participated at t=1. Thus, failing to account for the different future pathways between participation and outcome measurement periods overstates the effect of (early) participation.

Obviously, many factors other than borrowing dictate changes in consumption levels over time and, with a slight downturn in 2003, average consumption increased between v=1 and v=4 for both participants and non-participants, albeit at different pace. Specifically, except in the bad year 2003 in which case there was a consumption downturn, the pair wise causal effects, for the most part, overestimate impacts because the counterfactual paths are not taken into account.

Household consumption expenditure measurement period 1997 (v=1) 2000 (v=2) 2003 (v=3) 2006 (v=4) Timing of participation

Composite Pair wise Composite Pair wise Composite Pair wise Composite Pair wise

t=1 398.045 398.045*** (109.118) 388.733 529.716* (281.947) 859.674 371.162** (146.650) 1238.704 1487.966*** (546.147) t=2 133.156 133.156 (428.498) -568.872 -296.506 (263.564) 524.370 1457.280* (871.109) t=3 717.044 717.044* (462.315) -292.578 497.059 (1178.491) t=4 429.150 429.150 (1591.532)

Evidently, conventional parametric impact assessments that compare ‘ever participants’ to ‘never participants’ without considering the timing of the decision to participate and the different potential future pathways an individual household might have followed in the absence of the program would yield biased estimates.

Finally, given the relatively longer period the data set covers, including two drought years (1999/2000 and 2003) in between, it is interesting to see the implications of the effects of these differences in timings of participation on household consumption and hence relative capacity to cope with vulnerability during and after the drought years. Composite effects of participation in the first three periods i.e., at t = 1, 2, and 3, on annual consumption during the last three outcome measurement periods, i.e., v = 2, 3, and 4, are of interest here. Compared to controls, results suggest that the average annual consumption of the earliest (t=1) participants has increased steadily, including during and post drought years. Intuitively, sufficient time is needed for the cumulative impact of credit to take effect (King and Behrman, 2009). This is however not the case for later (t=2 and t=3) participants. In fact, although both participant groups have seen increased average consumption in the year they participated (which happened to be the drought years for both), in both cases, it has declined a year after participation (post drought years). A possible explanation for this is that households might have diverted loans to smooth consumption in the drought years, a common phenomenon despite DECSI’s claims of ‘productive’ use of credit. A study on the same MFI by Borchgrevink et al., (2005:68-69) finds indications of use of credit given for production purposes diverted to consumption during drought periods. This is also inline with the claim in chapter three that for households that are borrowing risk constrained, credit might be only useful as a last resort in times of distress. Moreover, the fact that loans are repaid after one year seems to explain the relative decline in participant households’ consumption in the post drought periods. Nevertheless, for the t = 2 participants, the result suggests this decline has been reversed in 20061. It can therefore be concluded that relative to non-participants, earlier participants gained better capacities to cope with shocks and the earlier the better. This conclusion has to be taken with caution though because the results explicitly compare variations of average consumption due to credit and not overall consumption variability due to shocks.

6.4.3 Effects of changes in the composition of treatment and control groups in